Detection of Multiple Eye Disorder using Deep Learning Techniques

Abstract

The global impact of vision impairment, affecting an astonishing 2.2 billion individuals, imposes a significant burden on public health, with a notable proportion of cases being preventable. Retinopathy, characterized by retinal vascular disease, stands as a leading cause of preventable blindness. Early detection is challenging due to asymptomatic early stages, necessitating the adoption of automated diagnostic approaches. Here, we used 3220 labelled fundus images from RFMID dataset to create a AI-powered neural platform that can identify several fundus diseases (39 classes). This study conducts an extensive review of deep learning architectures employed by researchers, encompassing models such as VGG-19, DenseNet-121, and Efficient Net- B0. This work introduces the development of a deep neural network model explicitly designed for classifying the RFMID dataset. Initial models displayed suboptimal accuracy, prompting efforts to reduce trainable parameters. Despite enhancements in computational efficiency, a notable increase in accuracy remained elusive. Subsequently, transfer learning was employed, involving model training on a diverse dataset before evaluating its performance on RFMID dataset. To address overfitting, dropout techniques were strategically applied, resulting in a final model showcasing improved accuracy.This study emphasizes how transfer learning and dropout strategies can improve the effectiveness and precision for diagnosing retinopathy. The consequences of these findings extend to the broader goal of leveraging advanced technologies to address global vision impairment and proactively prevent cases of avoidable blindness.